AI Science Tracker
EN

Vision models often encoded the same concept change differently across contexts

University of Tübingen; KAIST · 26 Oct 2026 · computer science

Discovery

Vision models often encoded the same concept change differently across contexts

Source check: Source checked. Not independent scientific confirmation.

Finding: Across 46 vision encoders, the direction for a change such as dry to wet was detectable within individual visual contexts much more often than across them: 97.8% versus 18.3% of the compared cases in that example. The contrast varied substantially by concept.

Technology: Pretrained vision encoders including ViT; learned embedding representations and fitted probes

Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

AI contribution
Learned vision embeddings supplied concept-transition vectors for within- and cross-context tests, and fitted probes tested whether aligned directions predicted transfer.
Test and result
The authors matched image embeddings for ten concept families, then compared within-context and cross-context direction consistency against random baselines. They also tested linear probe transfer across contexts.
Limitation / still unknown
Matching approximates changes in one concept using images that may differ in other ways. Detection rates depend on the chosen concepts, contexts, models, direction metric and statistical controls.
Evidence sources and annotation context

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

Evidence sources: https://arxiv.org/abs/2610.35857v1 · https://arxiv.org/html/2610.35857v1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 21 Nov 2026. Review scope: §§3–5, Table 1 and probe-transfer analysis.

No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.

Editorial assessment: Impact 2/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Open in the full catalog for filters, charts and related discoveries. Browse computer science discoveries.